TemPEST: Soft Template-Based Personalized EDM Subject Generation through Collaborative Summarization

Yu-Hsiu Chen, Pin-Yu Chen, Hong-Han Shuai, Wen-Chih Peng

Introduction

Electronic Direct Mail (EDM) remains a critical tool for e-commerce conversion, yet its effectiveness is often limited by generic messaging. While personalization is known to significantly boost open rates, generating tailored content at scale presents a unique challenge: balancing factual accuracy with individual user preferences. TemPEST (Soft Template-based Personalized EDM Subject Generator) addresses this by treating subject line generation as a collaborative summarization task, blending product descriptions with user-specific stylistic cues.

The Challenge of Personalized Marketing

Standard text summarization models prioritize "faithfulness" to the source text. However, in marketing, "attractiveness" is equally vital. Current AI solutions often fall short because they either lack the ability to adapt to diverse user tastes—ranging from those who prefer sensationalist "hooks" to those who favor straightforward descriptions—or they require labor-intensive manual labeling of product aspects. Furthermore, the "cold start" problem—where little is known about a new user's preferences—remains a persistent hurdle for recommendation-driven generation.

The TemPEST Architecture

TemPEST introduces a neural framework that integrates content summarization with collaborative filtering. The architecture is divided into three primary components:

  1. Retrieve and Re-rank: The system identifies successful historical subject lines to serve as "soft templates." Using Lucene for retrieval and Word2Vec for semantic refinement, it selects templates that provide a stylistic foundation without imposing rigid constraints.
  2. Dual Selective Encoders:
    • Template-aware Sequence Encoder (TSE): This module filters information between the source article and the template, ensuring the generated text remains grounded in product facts while adopting the template's style.
  3. User-aware Sequence Encoder (USE): This component acts as the personalization engine, analyzing a user’s click-through history to extract latent stylistic preferences and applying them to the article representation.
  4. Gated Attentive Autoencoder: To mitigate cold-start issues, TemPEST incorporates a collaborative filtering mechanism. By reconstructing missing user ratings through a fusion of item content and historical interactions, the model gains a deeper understanding of what a user finds appealing, which then guides the generation process.
The high-level architecture of TemPEST, illustrating the integration of content encoding, user history, and rating reconstruction.

Experimental Setup and the KKday Dataset

To evaluate the framework, the authors introduced KKday, a new benchmark dataset derived from a leading Asian travel e-commerce platform. The dataset comprises 17,617 products and over 278,000 click-through records from 21,379 users. This represents a significant contribution to the field, as it provides the scale necessary to train models on the intersection of long-form product descriptions (averaging 885 characters) and succinct, punchy subject lines (averaging 12.8 words).

Results and Impact

TemPEST was benchmarked against standard Sequence-to-Sequence (S2S) models and state-of-the-art template-based generators.

  • Linguistic Quality: TemPEST achieved a BLEU-1 score of 24.36, a substantial improvement over basic S2S models (20.59) and simple lead-sentence extraction (6.16).
  • User Preference: In a human evaluation study involving 81 participants, 61.4% of users preferred TemPEST-generated subjects over non-personalized alternatives.
  • Personalization Efficacy: Semantic analysis showed that TemPEST's outputs were 16% more aligned with a user's historical preferences than baseline models.
  • Recommendation Synergy: Notably, the joint task of generation and rating reconstruction improved recommendation metrics like Recall@10 and NDCG@10, suggesting that learning to "write" for a user also helps the system "understand" what products they might like.
Examples of personalized subject lines generated by TemPEST for the same product across different users, demonstrating stylistic adaptation.

Conclusion

TemPEST demonstrates that the "how" of marketing communication is just as important as the "what." By leveraging soft templates and collaborative filtering, the model successfully generates subject lines that are both faithful to the product and resonant with the individual user. While challenges remain regarding rare vocabulary and even finer-grained style control, TemPEST sets a new standard for personalized, automated marketing content.

Summary updated

Abstract

We address personalized Electronic Direct Mail (EDM) subject generation, which generates an attractive subject line for a product description according to user's preference on different contents or writing styles. Generating personalized EDM subjects has a few notable differences from generating text summaries. The subject has to be not only faithful to the description itself but also attractive to increase the click-through rate. Moreover, different users may have different preferences over the styles of topics. We propose a novel personalized EDM subject generation model named Soft Template-based Personalized EDM Subject Generator (TemPEST) to consider the aforementioned users' characteristics when generating subjects, which contains a soft template-based selective encoder network, a user rating encoder network, a summary decoder network and a rating decoder. Experimental results indicate that TemPEST is able to generate personalized topics and also effectively perform recommending rating reconstruction.
BibTeX
@inproceedings{chen2020tempest,
  author = {Yu-Hsiu Chen and Pin-Yu Chen and Hong-Han Shuai and Wen-Chih Peng},
  title = {{TemPEST: Soft Template-Based Personalized EDM Subject Generation through Collaborative Summarization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i05.6252},
}
Figure